Tested: AMD Ryzen AI 300 brings serious performance to Copilot+ PCs
Tested: AMD Ryzen AI 300 brings serious performance to Copilot+ PCs
We dive into the benchmarks to show how AMD's new Zen 5-based Ryzen AI 300 does against Intel's Core Ultra and Qualcomm's Snapdragon X Elite
You're viewing a single thread.
Is there a non-AI model?
I don’t want to pay money and watts for silicon I won’t use
It's the kind of thing that in a few years you'll be glad to have because some tasks will be offloaded from the GPU to the NPU, one I can think of is the background blur in video conferencing software or microphone noise suppression
GPUs are already really good at those, and GPUs are not a fad.
GPUs are already good at those things. But NPUs can be better at those things - and probably more energy efficient too - given optimized domain-specific models to work with.
Cpus are already good enough for terminal display, why would I want a GPU? ;)
This but unironically. I'm obsessed with off-grid, low power processing. I don't want a stupid fucking NPU I won't need nor use.
Then don't get one? I don't use tampons bit I don't feel the need to yell about not buying them
You can quite literally turn a GPU into an NPU by adding low-precision FP modes and then throwing away tons of silicon. Which isn't what AMD has traditionally done, and I doubt they'll go that way: They'll add the FP modes but keep the rest intact, that is, have no difference between GPU and tensor cores, so that all of the FLOPs can be used for either workload. They're also traditionally good at supporting multiple precision modes, FP16 is about exactly twice as fast as FP32 on their cards, very clean implementation. Haven't seen numbers yet (are some available?) but I abosutely wouldn't be surprised if FP8 is twice as fast, again.
Which brings me to "GPUs are really good at those things": With FP8, the same GPUs will be more than twice as good -- twice the FLOPS throughput, half the memory pressure. Let them advertise the stuff as AI whatever I don't care it's good and useful hardware.
That's still something that works more than good enough without the use of an NPU. We'll see if better use cases come out in future but I'd still prefer that power and die area go to raw CPU processing power.
You're likely already paying for silicon you don't use. Take AVX512 instructions (or other more exotic instructions) for example, or unused PCIe lanes, certain encoders on your GPU etc.
Also, people thinking they have no use for an NPU are almost certainly already using them, for example when taking a photo with their phone.
The closest thing is probably the 7840HS, 7940HS or the U variants (same thing at a lower TDP so slightly slower). The 8945HS, 8845HS, etc. are very similar but with a better NPU, which you won't want to pay for (and which still doesn't quite meet Copilot specs), so the 7xxx ones are better value. The performance of the AI 300 looks very similar to the 7940HS. Power efficiency may be better in the new one though.
It's actually extremely power efficient.